Free scoping read. A written assessment of your manuscript in two business days, at no cost.

Start one
Contents

Academy · Research Data Analytics & AI · Imprecision defined

Imprecision in Research: Definition and Assessment under GRADE

Imprecision does not concern whether a result is statistically significant. It concerns whether the confidence interval is narrow enough that both of its limits indicate the same decision. An interval spanning appreciable benefit and appreciable harm provides no answer, irrespective of the p-value, and an interval that excludes no effect may still be too wide to support action.

Pubrica AcademyWritten by the medical writing and biostatistics teams Reviewed byPubrica biostatistics team Published9 October 2026 Last checked9 October 2026 Reading time9 minutes Sources30 peer-reviewed

Imprecision is the least well handled of the GRADE domains, largely because it is conflated with significance testing. The question that GRADE poses is decision-theoretic: if the true effect were at the lower confidence limit, and separately if it were at the upper limit, would the same recommendation follow? If the answer differs at the two ends, the evidence is imprecise regardless of where the point estimate sits. GRADE provides explicit thresholds and sample-size triggers for making that judgement, and this article sets them out.

Learning objectives

  • What imprecision is, and how it differs from bias, inconsistency and indirectness
  • The GRADE thresholds for appreciable benefit and harm, and their application to a given interval
  • How optimal information size is calculated and what event count it implies at a conventional effect size
  • The explicit sample-size triggers GRADE gives for continuous outcomes and for very low event rates
  • What contextualised certainty assessment changed about imprecision rating
  • How to write an imprecision judgement that a reader can verify

Common misconceptions

  • “A significant result is precise enough.” An interval can exclude no effect and still be too wide to act on, because the magnitude it permits at one end implies a different decision from the magnitude at the other.
  • “A non-significant result means no effect.” It frequently means the study was too small to distinguish a worthwhile effect from no effect. This constitutes imprecision and should be rated as such rather than reported as evidence of absence.
  • “Imprecision is just small sample size.” Sample size is a principal determinant, but the judgement is about the interval relative to a decision threshold. A large trial with a low event rate can be imprecise, and a modest trial on a common outcome may not be.
  • “Imprecision is a kind of bias.” It is not. Bias moves the estimate away from the truth systematically; imprecision widens the range of truths compatible with the data. A precise estimate can be severely biased, and a wide interval can be entirely unbiased.
  • “There is no benchmark, so it is a matter of opinion.” GRADE offers explicit rough thresholds — a relative risk of 0.75 and 1.25 as a default for appreciable benefit or harm, a default appreciable relative risk reduction of 25%, and sample-size triggers of 400 total participants for continuous outcomes and 400 events for binary ones.

Definition of imprecision and its distinction from related concepts

Imprecision concerns the width of the confidence interval around an effect estimate, judged against the decision the estimate is meant to inform. It is one of the domains that can lower certainty in a body of evidence, and it is conceptually distinct from the others 6.

Not bias. Risk of bias asks whether the study design or conduct pushes the estimate systematically away from the truth. Imprecision asks how wide a range of truths is compatible with the data 4. An unbiased estimate from a small study is imprecise but not biased; a tight estimate from an unblinded trial with differential loss is precise but biased. Both reduce usable certainty, by different mechanisms and with different remedies — additional data reduce imprecision, whereas additional data do not correct bias.

Not inconsistency. Inconsistency is variation in the effect between studies. A set of studies can agree closely on an imprecise estimate, or disagree sharply while each is individually precise 6. Confusing the two leads to double-counting in a random-effects analysis, where between-study variation already widens the interval.

Not indirectness. Indirectness concerns whether the population, intervention, comparator or outcome matches the question. A perfectly precise estimate in the wrong population is a different problem.

The practical test is to specify the decision, state the effect at each end of the interval, and determine whether the decision differs between them. That is the entire judgement, and it is one that most discussion sections do not perform.

Thresholds proposed by GRADE

Judging whether an interval is too wide requires a view on what counts as an appreciable effect, and GRADE gives a rough default where no context-specific threshold exists: a relative risk of 0.75 at the lower end and 1.25 at the upper end, corresponding to a default appreciable relative risk reduction of 25% 4.

Applied to an interval, the rule is straightforward. If the confidence interval includes both a relative risk below 0.75 and one above 1.25, it spans appreciable benefit and appreciable harm, and certainty should be rated down for imprecision — often by two levels, since the evidence supports opposite recommendations at its two ends 4. If the interval sits entirely within 0.75 to 1.25, the evidence is reasonably taken as showing no appreciable effect rather than as uninformative, which is a different and often more useful conclusion 4.

Two cautions apply. First, these thresholds are a default for use when nothing better is available, not a standard. A mortality outcome may warrant a far smaller threshold for appreciable effect, and a minor symptom outcome a larger one; where a minimal important difference exists for the outcome, it should be used instead 1,4. Second, the thresholds apply to relative measures, so the absolute consequences at the baseline risk in the population of interest may be very different from what the relative interval suggests; this is the reason for presenting absolute effects alongside relative ones.

Optimal information size

The second way to judge imprecision is to ask whether the body of evidence contains as much information as a single adequately powered trial would have needed 4,27. That quantity is the optimal information size, and it is calculated with conventional error rates and a plausible effect size: an alpha of 0.05, a beta of 0.2, and a relative risk reduction in the range of 20% to 30% 4. At a relative risk reduction of 25%, the threshold works out at around 400 events 4.

The advantage of this approach is that it gives a check independent of the observed interval. A meta-analysis whose pooled interval looks narrow but which rests on far fewer than the required number of events reports a precision that the evidence does not support, typically because of a few small trials with extreme results, or because a random-effects model with poorly estimated between-study variance has produced an interval that is narrower than the evidence supports.

The underlying assumption should be noted. The 400-event figure depends on the 25% relative risk reduction that generated it. If the plausible effect of the intervention is 10%, the information required is substantially greater; if it is 50%, substantially less. The assumed relative risk reduction should be stated, because an optimal information size quoted without its assumption cannot be verified. See our pages on effect size in meta-analysis and statistical tests in meta-analysis.

400events meeting optimal information size at a 25% relative risk reduction
400total participants, trigger for continuous outcomes at an effect size of 0.2 SD
4,000participants for prognostic balance at a 5% baseline risk

Three GRADE sample-size triggers for imprecision. The event figure is calculated with an alpha of 0.05, a beta of 0.2 and a relative risk reduction in the 20% to 30% range; it is a function of the assumed reduction and should always be quoted with it.

Continuous outcomes and very low event rates

GRADE gives explicit sample-size triggers for two situations where interval width alone is a poor guide.

Continuous outcomes. Rating down for imprecision should be considered when the total sample size across studies is below 400, a figure derived from the number required to detect an effect size of 0.2 standard deviations with conventional error rates 4. Below that, the interval around a standardised mean difference is rarely narrow enough to distinguish a small worthwhile effect from nothing — see our page on standardised mean difference estimation for what that measure does and does not convey 4.

Very low event rates. Where events are very rare, the relative measure can look precise while resting on a handful of events 4. GRADE gives sample-size triggers in this situation of 2,000 participants as a minimum and 4,000 at the higher end, reflecting that very low event rates require large populations before any estimate is informative.

Prognostic balance. A related trigger concerns whether randomisation is likely to have balanced prognostic factors: at a baseline risk of 5%, around 4,000 participants are needed for that assumption to be reasonable. Below this number, chance imbalance remains a real possibility, and the uncertainty this introduces is not captured by the confidence interval at all.

GRADE triggers for considering a rating down for imprecision

SituationTriggerWhere it comes from
Binary outcome, relative measureConfidence interval includes a relative risk below 0.75 and above 1.25Default threshold for appreciable benefit or harm, equivalent to a 25% relative risk reduction
Binary outcome, information sizeFewer than about 400 events accumulatedOptimal information size at alpha 0.05, beta 0.2 and a 25% relative risk reduction
Continuous outcomeTotal sample size across studies below 400Number required to detect an effect size of 0.2 standard deviations at conventional error rates
Very low event rateBelow 2,000 participants as a minimum, and up to 4,000 at the higher endExplicit GRADE sample-size triggers for very low event rates
Prognostic balance at 5% baseline riskAround 4,000 participantsNumber at which randomisation can comfortably be assumed to have balanced prognostic factors

Thresholds and triggers as set out in GRADE guidance on imprecision. The optimal information size figure depends on the assumed relative risk reduction, which GRADE places in the range of 20% to 30%; 400 events corresponds to the 25% assumption.

Contextualised certainty assessment

GRADE guidance 35, published in 2022, updated how imprecision is rated for contextualised certainty assessments 2. The underlying distinction is between asking whether there is any effect at all and asking whether the effect is large enough to matter in a specific decision context; the two questions give different imprecision ratings on the same data.

In a non-contextualised assessment, the question is whether the evidence shows an effect, and the relevant comparison is between the interval and the null 1. In a contextualised assessment, the question is whether the effect crosses a decision threshold, and the relevant comparison is between the interval and that threshold. An interval that comfortably excludes no effect may still straddle the threshold at which a guideline panel would change its recommendation, in which case it is imprecise for the contextualised question and precise for the other 2.

The practical requirement this creates is one of disclosure. Authors should state the question against which imprecision is rated and, if it is the contextualised question, the threshold used and its source. A rating made against an unstated threshold cannot be checked or reproduced, which is the same objection that applies to an optimal information size quoted without its assumed effect size.

Reporting a verifiable imprecision judgement

Four elements are required. Each is brief, and all four are routinely omitted.

The interval, with the absolute effect. Both relative and absolute, because the same relative interval has very different consequences at different baseline risks, and the decision is made on absolute terms.

The threshold used, and its source. A minimal important difference for the outcome if one exists, or the GRADE default of a relative risk of 0.75 to 1.25 with the reason for relying on it 4.

The comparison. Whether the interval crosses the threshold, and in which direction. This is the decisive sentence, and it should state explicitly whether the two ends of the interval imply the same decision.

The information size, where relevant. The number of events or participants accumulated against the number needed, with the assumed relative risk reduction stated. For a continuous outcome, the total sample size against the 400 trigger 4.

A judgement containing these four elements permits productive disagreement, which is its purpose. A judgement stating only that the evidence was downgraded for imprecision does not. Our statistical review service checks this kind of judgement, and the PRISMA reporting items cover where it belongs in a review.

The imprecision sentence in a review’s certainty assessment (figures below are illustrative, to show the form of the judgement)

Unverifiable downgrade

The certainty of evidence for this outcome was rated as low. Although the pooled estimate favoured the intervention, the result did not reach statistical significance and the confidence interval was wide, so the evidence was downgraded for imprecision. Further large trials are needed.

Significance is used as the criterion. Imprecision is about whether both ends of the interval support the same decision, not about whether the interval excludes the null.

No interval given. “Wide” is not assessable; a reader cannot check a judgement whose input is not reported.

No absolute effect. The decision is made in absolute terms, and the same relative interval means different things at different baseline risks.

No threshold. Without a stated threshold for appreciable effect, there is nothing against which the interval was compared.

No information size. Event count against the required number is the independent check, and it is absent.

No reason for the number of levels. One level or two is a substantive judgement, and “downgraded for imprecision” conceals it.

Checkable judgement

The pooled relative risk was 0.82 with a 95% confidence interval of 0.61 to 1.10, corresponding at the control-arm risk of 12% to an absolute difference of 22 fewer events per 1,000 (95% CI 47 fewer to 12 more). No minimal important difference has been established for this outcome, so we applied the GRADE default thresholds for appreciable effect of a relative risk of 0.75 and 1.25. The interval includes a relative risk of 0.61, which is an appreciable benefit, and 1.10, which is neither appreciable benefit nor appreciable harm; it therefore crosses the lower threshold but not the upper one, and a recommendation at the lower limit would differ from one at the upper limit. The analysis accumulated 186 events against an optimal information size of approximately 400 events, calculated at an alpha of 0.05, a beta of 0.2 and an assumed relative risk reduction of 25%. On both grounds we rated down one level for imprecision; we did not rate down two levels, because the interval does not extend to appreciable harm. Certainty was rated moderate, with imprecision the only domain downgraded.

The relative and absolute intervals are both given, so the decision-relevant magnitude is visible.

The threshold is stated with its source and with the reason a default was used, which makes the comparison reproducible.

The crossing is described explicitly at both ends, including the conclusion that the two ends imply different recommendations.

The information size check is given with the assumed relative risk reduction, so the 400-event figure can be recalculated under a different assumption.

The number of levels is justified against the specific feature of the interval that decides it, which is where most disagreement between reviewers lies.

A reader who disagrees with the threshold can redo the judgement from the reported numbers, which is the purpose of reporting it.

The first version asks the reader to accept a conclusion; the second allows the reader to recompute it. Because the choice of threshold is the element most open to legitimate disagreement, a judgement that conceals the threshold conceals the element that most warrants scrutiny.

Key points

Key points

Imprecision asks whether the confidence interval is narrow enough that both of its ends point to the same decision — not whether the result is significant. Where no outcome-specific minimal important difference exists, GRADE offers rough default thresholds for appreciable effect of a relative risk of 0.75 and 1.25, equivalent to a 25% relative risk reduction; an interval spanning both is a candidate for rating down two levels, since it supports opposite recommendations at its ends. A second, independent check is the optimal information size: calculated at an alpha of 0.05, a beta of 0.2 and a relative risk reduction in the 20% to 30% range, it comes to around 400 events at a 25% reduction. For continuous outcomes, rating down should be considered below a total of 400 participants, derived from the number needed to detect an effect size of 0.2 standard deviations; for very low event rates, triggers of 2,000 and 4,000 participants apply, and around 4,000 are needed for prognostic balance at a 5% baseline risk. GRADE guidance 35, published in 2022, updated imprecision rating for contextualised certainty assessments; authors should therefore state the question against which certainty is rated and the threshold used.

Application to your study

Application to your study

  1. The decision should be specified before the interval is judged. The recommendation that would follow if the truth lay at each limit should be recorded; if the two differ, the evidence is imprecise irrespective of the p-value.
  2. An outcome-specific threshold should be used where one exists. A minimal important difference for the outcome is preferable to the GRADE default of 0.75 to 1.25, which is a fallback rather than a standard.
  3. Absolute as well as relative effects should be reported. The same relative interval implies very different absolute consequences at different baseline risks, and decisions are made in absolute terms.
  4. Information size should be checked independently. Events should be counted against the required number, and the assumed relative risk reduction should be stated, since the 400-event figure is a function of the 25% assumption.
  5. The continuous-outcome trigger should be applied. Rating down should be considered below 400 total participants, and it should be established whether the standardised mean difference interval can distinguish 0.2 standard deviations from zero.
  6. The question against which certainty was rated should be stated. The rating may be non-contextualised, comparing the interval with the null, or contextualised, comparing it with a decision threshold; if contextualised, the threshold and its source should be given.
  7. A non-significant, imprecise result should not be reported as evidence of no effect. The accurate statement is that the study could not distinguish a worthwhile effect from no effect.

Frequently asked questions

They are different failures with different remedies. Bias is systematic: the design or conduct of a study pushes the estimate away from the truth in a particular direction, and collecting more data under the same design makes the biased estimate more precise without making it more correct. Imprecision is random: the estimate is compatible with a wide range of true values, and more data narrows that range. An unbiased small study is imprecise but not biased; a large unblinded trial with differential dropout can be precise and seriously biased. Both reduce the certainty that can be placed in a body of evidence, which is why GRADE rates them as separate domains rather than merging them.

It depends on what constitutes an appreciable effect for the outcome, which is why the question has no universal numeric answer. Where a minimal important difference exists for the outcome, it should be used. Where none exists, GRADE offers rough default thresholds of a relative risk of 0.75 and 1.25, corresponding to a default appreciable relative risk reduction of 25%. An interval including both an effect below 0.75 and one above 1.25 spans appreciable benefit and appreciable harm, supports opposite recommendations at its two ends, and is a candidate for rating down two levels. An interval sitting entirely between 0.75 and 1.25 is better read as showing no appreciable effect than as uninformative.

It is the amount of information a body of evidence would need in order to be as informative as a single adequately powered trial, and it provides a check on precision that does not depend on the observed interval. It is calculated with conventional error rates — an alpha of 0.05 and a beta of 0.2 — together with a plausible relative risk reduction, which GRADE places in the range of 20% to 30%. At a relative risk reduction of 25%, the threshold is around 400 events. The figure is only as reliable as the assumed reduction: at a genuinely plausible 10% reduction, substantially more information is needed. The assumption should always be quoted alongside the number.

Usually it means the study could not tell the difference between a worthwhile effect and none, which is a statement about the study rather than about the intervention. The two cases are distinguished by examining the interval against a threshold for appreciable effect. If the interval lies entirely within the no-appreciable-effect zone — for the GRADE default, between a relative risk of 0.75 and 1.25 — then concluding no appreciable effect is reasonable. If it extends beyond either threshold, the correct conclusion is that the evidence is imprecise and does not settle the question. Reporting the second case as evidence of absence is one of the most consequential errors in the literature.

Against the sample-size trigger as well as the interval. GRADE suggests considering a rating down when the total sample size across studies is below 400, a figure derived from the number of participants needed to detect an effect size of 0.2 standard deviations at conventional error rates. Below that, the interval around a standardised mean difference is rarely able to distinguish a small but worthwhile effect from nothing. Alongside the count, the same decision test as for binary outcomes applies: the interval should be expressed in the original units of the instrument where possible, both ends should be compared with the minimal important difference for that instrument, and it should be stated whether the two ends imply the same recommendation.

GRADE guidance 35, published in 2022, updated imprecision rating for contextualised assessments. The distinction is between two different questions: whether there is any effect at all, and whether the effect is large enough to matter in a specific decision. In a non-contextualised assessment the interval is compared with the null; in a contextualised assessment it is compared with a decision threshold. The same data can therefore give different imprecision ratings depending on which question is being asked; an interval that excludes no effect may still straddle the threshold at which a panel would change its recommendation. The practical requirement is disclosure: authors should state the question against which certainty was rated and, if contextualised, the threshold and its source.

Because precision is driven by the number of events rather than the number of participants. A trial of 5,000 people with 12 events contains about as much information about a relative effect as a trial of 200 people with 12 events. GRADE gives explicit triggers for this situation, suggesting sample sizes of 2,000 participants as a minimum and up to 4,000 at the higher end where event rates are very low. A related trigger concerns whether randomisation can be assumed to have balanced prognostic factors: at a baseline risk of 5%, around 4,000 participants are needed for that assumption to be reasonable, and the extra uncertainty from possible chance imbalance below that is not reflected in the confidence interval at all.

References

  1. Zeng L, Brignardello-Petersen R, Hultcrantz M, Mustafa RA, Murad MH, Iorio A, et al. GRADE Guidance 34: update on rating imprecision using a minimally contextualized approach. Journal of Clinical Epidemiology. 2022;150:216-224. doi:10.1016/j.jclinepi.2022.07.014 read 9 October 2026. doi.org/10.1016/j.jclinepi.2022.07.014
  2. Schunemann HJ, Neumann I, Hultcrantz M, Brignardello-Petersen R, Zeng L, Murad MH, et al. GRADE guidance 35: update on rating imprecision for assessing contextualized certainty of evidence and making decisions. Journal of Clinical Epidemiology. 2022;150:225-242. doi:10.1016/j.jclinepi.2022.07.015 read 9 October 2026. doi.org/10.1016/j.jclinepi.2022.07.015
  3. Wan F, Colditz GA, Sutcliffe S. Matched Versus Unmatched Analysis of Matched Case-Control Studies. American Journal of Epidemiology. 2021;190(9):1859-1866. doi:10.1093/aje/kwab056 read 9 October 2026. doi.org/10.1093/aje/kwab056
  4. Guyatt GH, Oxman AD, Kunz R, Brozek J, Alonso-Coello P, Rind D, et al. GRADE guidelines 6. Rating the quality of evidence - imprecision. Journal of Clinical Epidemiology. 2011;64(12):1283-1293. doi:10.1016/j.jclinepi.2011.01.012 read 9 October 2026. doi.org/10.1016/j.jclinepi.2011.01.012
  5. Guyatt G, Oxman AD, Akl EA, Kunz R, Vist G, Brozek J, et al. GRADE guidelines: 1. Introduction - GRADE evidence profiles and summary of findings tables. Journal of Clinical Epidemiology. 2011;64(4):383-394. doi:10.1016/j.jclinepi.2010.04.026 read 9 October 2026. doi.org/10.1016/j.jclinepi.2010.04.026
  6. Balshem H, Helfand M, Schunemann HJ, Oxman AD, Kunz R, Brozek J, et al. GRADE guidelines: 3. Rating the quality of evidence. Journal of Clinical Epidemiology. 2011;64(4):401-406. doi:10.1016/j.jclinepi.2010.07.015 read 9 October 2026. doi.org/10.1016/j.jclinepi.2010.07.015
  7. Schulz KF, Altman DG, Moher D. CONSORT 2010 Statement: updated guidelines for reporting parallel group randomised trials. Trials. 2010;11(1):32. doi:10.1186/1745-6215-11-32 read 9 October 2026. doi.org/10.1186/1745-6215-11-32
  8. Bassler D, Briel M, Montori VM, Lane M, Glasziou P, Zhou Q, et al. Stopping randomized trials early for benefit and estimation of treatment effects: systematic review and meta-regression analysis. JAMA. 2010;303:1180-1187. read 9 October 2026. school16.sipta.org
  9. Guyatt GH, Oxman AD, Vist GE, Kunz R, Falck-Ytter Y, Alonso-Coello P, et al. GRADE: an emerging consensus on rating quality of evidence and strength of recommendations. BMJ. 2008;336:924-926. read 9 October 2026. cobe.paginas.ufsc.br
  10. Schunemann HJ, Oxman AD, Brozek J, Glasziou P, Jaeschke R, Vist GE, et al. Grading quality of evidence and strength of recommendations for diagnostic tests and strategies. BMJ. 2008;336:1106-1110. read 9 October 2026. cobe.paginas.ufsc.br
  11. Guyatt GH, Oxman AD, Kunz R, Vist GE, Falck-Ytter Y, Schunemann HJ. What is "quality of evidence" and why is it important to clinicians? BMJ. 2008;336:995-998. read 9 October 2026. cobe.paginas.ufsc.br
  12. von Elm E, Altman DG, Egger M, Pocock SJ, Gotzsche PC, Vandenbroucke JP; STROBE Initiative. Strengthening the reporting of observational studies in epidemiology (STROBE) statement: guidelines for reporting observational studies. BMJ. 2007;335(7624):806-808. doi:10.1136/bmj.39335.541782.AD read 9 October 2026. doi.org/10.1136/bmj.39335.541782.AD
  13. von Elm E, Altman DG, Egger M, Pocock SJ, Gotzsche PC, Vandenbroucke JP; STROBE Initiative. The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies. Epidemiology. 2007;18(6):800-804. doi:10.1097/EDE.0b013e3181577654 read 9 October 2026. doi.org/10.1097/EDE.0b013e3181577654
  14. Ridker PM, Cook NR, Lee I-M, Gordon D, Walter JM, Buring JE, et al. A randomized trial of low-dose aspirin in the primary prevention of cardiovascular disease in women. New England Journal of Medicine. 2005;352(13):1293-1304. doi:10.1056/NEJMoa050613 read 9 October 2026. doi.org/10.1056/NEJMoa050613
  15. Ioannidis JP. Contradicted and initially stronger effects in highly cited clinical research. JAMA. 2005;294(2):218-228. doi:10.1001/jama.294.2.218 read 9 October 2026. doi.org/10.1001/jama.294.2.218
  16. Montori VM, Devereaux PJ, Adhikari NK, Burns KEA, Eggert CH, Briel M, et al. Randomized trials stopped early for benefit: a systematic review. JAMA. 2005;294:2203-2209. read 9 October 2026. school16.sipta.org
  17. Lawlor DA, Davey Smith G, Ebrahim S. Commentary: the hormone replacement-coronary heart disease conundrum: is this the death of observational epidemiology? International Journal of Epidemiology. 2004;33(3):464-467. doi:10.1093/ije/dyh124 read 9 October 2026. doi.org/10.1093/ije/dyh124
  18. Doll R, Peto R, Boreham J, Sutherland I. Mortality in relation to smoking: 50 years' observations on male British doctors. BMJ. 2004;328(7455):1519. doi:10.1136/bmj.38142.554479.AE read 9 October 2026. doi.org/10.1136/bmj.38142.554479.AE
  19. Schulz KF, Grimes DA. Case-control studies: research in reverse. The Lancet. 2002;359(9304):431-434. doi:10.1016/S0140-6736(02)07605-5 read 9 October 2026. doi.org/10.1016/S0140-6736(02)07605-5
  20. Grimes DA, Schulz KF. Cohort studies: marching towards outcomes. The Lancet. 2002;359:341-345. doi:10.1016/S0140-6736(02)07500-1 read 9 October 2026. doi.org/10.1016/S0140-6736(02)07500-1
  21. King G, Zeng L. Estimating risk and rate levels, ratios and differences in case-control studies. Statistics in Medicine. 2002;21(10):1409-1427. doi:10.1002/sim.1032 read 9 October 2026. doi.org/10.1002/sim.1032
  22. Willan AR. Analysis, sample size, and power for estimating incremental net health benefit from clinical trial data. Controlled Clinical Trials. 2001;22:228-237. read 9 October 2026. ouci.dntb.gov.ua
  23. Beral V, Hermon C, Kay C, Hannaford P, Darby S, Reeves G. Mortality associated with oral contraceptive use: 25 year follow up of cohort of 46 000 women from Royal College of General Practitioners' oral contraception study. BMJ. 1999;318:96-100. read 9 October 2026. qmplus.qmul.ac.uk
  24. Willan AR, O'Brien BJ. Sample size and power issues in estimating incremental cost-effectiveness ratios from clinical trials data. Health Economics. 1999;8:203-211. read 9 October 2026. ouci.dntb.gov.ua
  25. Briggs AH, Gray AM. Power and sample size calculations for stochastic cost-effectiveness analysis. Medical Decision Making. 1998;18(Suppl):S81-S92. read 9 October 2026. ouci.dntb.gov.ua
  26. Rothman N, Cantor KP, Blair A, Bush D, Brock JW, Helzlsouer K, et al. A nested case-control study of non-Hodgkin lymphoma and serum organochlorine residues. The Lancet. 1997;350:240-244. read 9 October 2026. qmplus.qmul.ac.uk
  27. Pogue JM, Yusuf S. Cumulating evidence from randomized trials: utilizing sequential monitoring boundaries for cumulative meta-analysis. Controlled Clinical Trials. 1997;18:580-593. read 9 October 2026. school16.sipta.org
  28. Bland JM, Altman DG. Regression towards the mean. BMJ. 1994;308:1499. read 9 October 2026. qmplus.qmul.ac.uk
  29. Pocock SJ, Hughes MD. Practical problems in interim analyses, with particular regard to estimation. Controlled Clinical Trials. 1989;10:209S-221S. read 9 October 2026. school16.sipta.org
  30. Doll R, Hill AB. Smoking and carcinoma of the lung; preliminary report. British Medical Journal. 1950;2(4682):739-748. doi:10.1136/bmj.2.4682.739 read 9 October 2026. doi.org/10.1136/bmj.2.4682.739

Every source above is a peer-reviewed article, a publisher’s record of one, or a standard issued by the body that owns it. Each carries the date it was last read. Where a standard moves, this article is revised rather than left up with a note.

How to cite this article

Pick your referencing style and copy it. Your university or target journal may vary the punctuation — check its guide if it has one.

Pubrica Academy. (2026, October 9). Imprecision in Research: Definition and Assessment under GRADE. Pubrica. https://pubrica.com/academy/concepts-definitions/what-is-imprecision-in-research/

This article is written to be learned from. Cite it if you are referring to what it says; if you are using the evidence behind it, cite the sources listed above instead.

Imprecision judgement worksheet

One block should be completed for each outcome. Completing it produces the four or five sentences that belong in the certainty assessment and makes the threshold explicit, so that a reader can evaluate it.

IMPRECISION JUDGEMENT - ONE PER OUTCOME

Outcome: ____________________________________________________
Type: [ ] binary  [ ] continuous  [ ] time to event

1. THE ESTIMATE
   Relative measure and 95% CI: ________________________________
   Control-arm / baseline risk used: ______%
   Absolute effect per 1,000, with CI: ________________________
   Instrument and units (continuous outcomes): ________________

2. THE THRESHOLD FOR AN APPRECIABLE EFFECT
   [ ] Outcome-specific minimal important difference
       Value: ______  Source: ______________________
   [ ] GRADE default (RR 0.75 and 1.25; 25% RRR)
       Reason for using the default: ___________________________

3. THE DECISION TEST
   If the truth were at the LOWER limit, we would recommend:
     ____________________________________________________
   If the truth were at the UPPER limit, we would recommend:
     ____________________________________________________
   Are these the same?  [ ] yes  [ ] no
   Interval crosses: [ ] lower threshold  [ ] upper threshold
                     [ ] both  [ ] neither

4. INFORMATION SIZE
   Binary outcomes:
     Events accumulated: ______
     Optimal information size: ______ events
     Assumed RRR used in that calculation: ______%
     (alpha ______  beta ______ )
   Continuous outcomes:
     Total participants: ______   Trigger: 400
   Very low event rate:
     Participants: ______   Triggers: 2,000 minimum / 4,000 higher
   Prognostic balance:
     Baseline risk ______%  Participants ______
     (around 4,000 needed at a 5% baseline risk)

5. WHICH QUESTION ARE YOU RATING AGAINST?
   [ ] Non-contextualised - interval versus the null
   [ ] Contextualised - interval versus a decision threshold
       Threshold used: ______  Source: ______________________

6. RATING
   Rate down for imprecision?  [ ] no  [ ] one level  [ ] two levels
   Reason, naming the specific feature of the interval that decides it:
     ____________________________________________________
   Other domains downgraded for this outcome: ________________
   Final certainty: [ ] high [ ] moderate [ ] low [ ] very low

SENTENCES TO AVOID
  [ ] "did not reach statistical significance" as the imprecision reason
  [ ] "the confidence interval was wide" without giving it
  [ ] "no effect was found" when the interval crosses a threshold
  [ ] "downgraded for imprecision" with no threshold stated

Pubrica AcademyWritten and maintained by Pubrica’s subject matter experts. Articles are revised when the underlying standards change, and each article states the date on which it was last reviewed.

How Pubrica can help

Pubrica writes imprecision judgements that can be checked: the relative and absolute interval, the threshold used and its source, an explicit statement of whether the two ends of the interval imply the same decision, and the event count against the optimal information size with the assumed effect size shown. Where a review has downgraded for imprecision without stating a threshold, Pubrica identifies the omission before peer review, and where a non-significant result has been reported as evidence of no effect, Pubrica revises the statement to reflect what the data support.

Find something else

Search the Pubrica Academy

Access is free and requires no account. Enter a topic, or open the full library to filter by category, sub-topic or resource type.

Pubrica reviews evidence profiles and identifies the imprecision ratings that peer reviewers are likely to question.

A scoping assessment is free and takes two business days. It provides a written assessment of what the work requires and which specialist would undertake it.

Get a free quote

Regulations

Check your project against this rule

  1. 1Your work
  2. 2Your details
  3. 3Send
What do you need?

Material you send is held under signed confidentiality, with access limited to the assigned team and encrypted transfer. It is not used to train anything and is not passed to third-party tools. Ask for the agreement before you send anything you would rather cover first.